Install

git clone https://github.com/slima4/agent-message
cd agent-message
./install.sh

Idempotent — safe to re-run. Open a new terminal after first install so the shell function loads.

What gets installed

Path	Purpose
`~/.claude/commands/message-{send,inbox,reply}.md`	Claude Code slash-command prompts
`~/.agent-message-cmd`	Python wrapper — single entry point used by the slash commands and any other agent
`~/.agent-message.sh`	`msg` shell function, sourced from `~/.zshrc` and `~/.bashrc` via an idempotent `# >>> agent-message >>>` block
`${XDG_STATE_HOME:-~/.local/state}/agent-message/`	Default shared message directory (`AGENT_MESSAGE_DIR` overrides)

Flags

Flag	Default	Effect
`--dir <path>`	`${XDG_STATE_HOME:-~/.local/state}/agent-message`	Override message dir
`--commands <path>`	`~/.claude/commands`	Override Claude commands dir
`--shell <path>`	`~/.agent-message.sh`	Override shell helper install path
`--bin <path>`	`~/.agent-message-cmd`	Override wrapper install path
`--no-shell`	install shell	Skip shell helper
`--uninstall`	install	Remove everything

Integrations

Wire up other agent tools with one flag. Global integrations install once and cover every repo; per-repo integrations target the current directory.

Flag	Scope	Writes
`--integrate=cursor`	global	`~/.cursor/rules/agent-message.mdc`
`--integrate=copilot-cli`	global	`~/.copilot/copilot-instructions.md`
`--integrate=antigravity`	global	`~/.gemini/AGENTS.md` (Antigravity + Gemini CLI)
`--integrate=codex`	global	`~/.codex/AGENTS.md` + sandbox writable root in `~/.codex/config.toml`
`--integrate=copilot`	per-repo	`.github/copilot-instructions.md` (Copilot Chat)
`--integrate=antigravity-repo`	per-repo	`./AGENTS.md` (cross-tool, opt-in)
`--integrate=zed`	per-repo	`./.rules`
`--integrate=all`	mixed	every flag above except `antigravity-repo`
`--integrate=auto`	mixed	detect installed tools and integrate them
`--integrate=select`	mixed	interactive menu of every tool with its state

auto only runs the global integrations — a global signal ("you have Zed installed") says nothing about whether the current directory is the repo where per-repo rules belong. Per-tool guides: integrations/.

Every ./install.sh reports integration state, so nothing installs silently half-connected. Wired tools appear as an agents: row inside the main block; anything detected but unwired gets its own paragraph with the command to run:

  commands: ~/.claude/commands/{message-send,message-inbox,message-reply}.md
  wrapper:  ~/.agent-message-cmd
  dir:      ~/.local/state/agent-message  (per-agent logs: log-<alias>.jsonl)
  agents:   antigravity, codex

Not wired yet (detected on this machine):
  cursor       → ./install.sh --integrate=cursor

  all at once: ./install.sh --integrate=auto
  or pick from a menu: ./install.sh --integrate=select

--integrate=select opens a menu instead. It is the only mode that offers the per-repo writers, because you can see which directory they would write to before confirming:

agent-message integrations   (cwd: /Users/you/dev/my-web)

  [1] cursor            global       detected
  [2] copilot-cli       global       not found
  [3] antigravity       global       WIRED
  [4] codex             global       detected
  [5] copilot           ./my-web     detected
  [6] antigravity-repo  ./my-web     detected
  [7] zed               ./my-web     not found

Numbers to pick (space-separated), a=all detected global, Enter=none, q=quit
> 4 5

a picks the detected global tools only — same set as auto. Per-repo writers need an explicit number. The menu needs a terminal; piped into curl … | bash it prints a notice and falls back to auto. It also works with --uninstall, where wired tools are marked WIRED.

Requirements

python3 — preinstalled on macOS, every Linux distro. The installer pre-flights and refuses if missing.
A POSIX shell (bash or zsh).
That's it. No pip, no npm, no Docker.

Permission tip (Claude Code)

Claude Code's safety detector flags Python f-strings inside heredocs as "expansion obfuscation" and prompts for approval on every send. To skip the prompt without granting blanket python3 access, allowlist only the wrapper:

{ "permissions": { "allow": ["Bash(/Users/<you>/.agent-message-cmd:*)"] } }

Add it to ~/.claude/settings.json. The rule allows ONLY the wrapper, nothing else.

Verifying / contributing

If you cloned the repo to hack on it:

./test.sh        # 51 tests, pure bash + python3, no other deps

CI runs the same suite on Ubuntu and macOS plus shellcheck on every push and PR. See CONTRIBUTING.md for line budgets, the single-writer invariant, and the smaller/cheaper/faster rule.

Uninstall

./install.sh --uninstall

Removes the slash commands, the wrapper, the shell helper, the per-agent logs + caches in the message dir, and the rc-block from ~/.zshrc / ~/.bashrc. Does not touch .agent-message files in your repos.